AI engines build each answer fresh, so the sources they cite change from day to day. Ahrefs followed about 37,500 citations for 20 days and found 50% to 83% were gone the next day, depending on the engine. About 75% of the lost ones came back. One missing citation is usually noise, so judge the pattern over weeks.
How often do AI citations disappear?
Often. On every engine Ahrefs tested, at least half of the day-one citations were gone the next day. In Ahrefs' study (Oct 29, 2026), five AI tools got the same questions daily for 20 days, and Ahrefs followed every website cited on day one.
| AI tool | Gone by day two | Lasted all 20 days |
|---|---|---|
| Google AI Overviews | 50% | 6.6% |
| Google AI Mode | 58% | 6.8% |
| ChatGPT | 59% | 4.0% |
| Gemini | 81% | 0.2% |
| Copilot | 83% | 0.5% |
Gemini and Copilot look worse partly because so many of their answers cited nobody. In Ahrefs' data, 64% of Gemini's answers and 53% of Copilot's had no citations at all. Skip those empty answers and their next-day loss falls to 58%, about the same as ChatGPT and AI Mode.
Ahrefs sells an AI citation tracker (Brand Radar), so keep that in mind. SISTRIX's longer study below, run by a different company, points the same way.
Do lost AI citations come back?
Most do. About 75% of the citations that dropped out at least once came back within the 20 days, and about half of those were back within three days. Only 20% to 27% of the websites cited on day one dropped out and never returned.
How long a citation had already lasted predicted whether it would return:
| Lasted before first going missing | Never came back: AI Overviews | AI Mode | ChatGPT |
|---|---|---|---|
| 1 day | 34% | 41% | 35% |
| 2 to 7 days | 8% | 10% | 4% |
| 8 to 14 days | 4% | 2% | 0.7% |
So a brand new citation is the least safe kind. Don't count it as a win until it has shown up for several days. And a citation that held for a week and then dropped will very likely come back without you doing anything.
Why would a citation vanish when nothing changed on your page?
Usually the engine changed its picks. Before it answers, an AI engine runs its own searches (fan-out searches), and those change from run to run. Ahrefs found ChatGPT ran exactly the same searches as the day before only 26% of the time.
Changing searches explain only part of it. Even when ChatGPT's searches stayed the same, the websites it cited still changed 81% of the time, against 91% when the searches changed. Ahrefs reads that as a sign that much of the churn happens after the search, when the model picks which pages to cite.
Two other things look like losses and aren't:
- The engine cited nobody. Common on Gemini and Copilot, as above.
- The engine dropped your link but kept your name. On Google's tools and ChatGPT, 12% to 16% of brands lost the citation the next day but were still named in the answer. A citation and a mention are two different counts, which is why we count mentions and citations separately.
Are some AI engines more stable than others?
Yes. Google's AI Overviews hold on to their sources best. They had the fewest next-day drops in Ahrefs' data, and in SISTRIX's data no source changed at all in 53% of prompts. Among the other engines, the order depends on the study, and each study measures something different.
| Study | What it measured | What it found |
|---|---|---|
| Ahrefs, Oct 2026, 20 days | Websites cited day to day | Gone by day two: AI Overviews 50%, ChatGPT 59%, Copilot 83% |
| SISTRIX, 17 weeks to Apr 2026, 82,619 prompts in six countries | Cited domains week to week | Weekly churn: AI Mode 56%, ChatGPT 74%. No AI Overview source changed in 53% of prompts |
| Chen and Koudas, Oct 2026, two 90-day collections | Which of a site's pages get cited, across 14-day windows | Still shared two weeks later: Gemini 47%, ChatGPT 40%, Perplexity 22%. Ten weeks later: 29%, 14%, 0% |
SISTRIX describes AI Mode answers as a fixed core plus a carousel. In 86.5% of prompts a core of 1 to 5 domains stayed put, while 89% of the other domains rotated every week. If you're in the core, you tend to stay. If you're in the carousel, you come and go.
SISTRIX also sells a prompt tracking tool. None of these studies covered Claude, Grok or Muse, so there's no public data yet on how stable their citations are.
Which citations tend to stick?
Pages that already rank in Google, pages whose title matches the question, and a brand's own site when the question is a fact about that brand. These are correlations from Ahrefs' data, not proof of cause.
- Ranking in Google. On AI Overviews and AI Mode, pages that ranked anywhere in Google for the question stayed cited for a week or more 28% to 29% of the time, against 4% for pages that didn't rank. On ChatGPT the gap was smaller (11% vs 4%). Ahrefs notes this part covered only 230 questions. More on how a ranking carries over in does ranking on Google get you cited in AI answers.
- Search traffic. On AI Overviews, pages cited for 15 days or more had a median of 88 estimated monthly organic visits, against 8 for pages cited for one day.
- A title that matches the question. On AI Overviews and AI Mode, the typical match score between a page's title and description and the question rose from 0.54 for one-day citations to 0.65 to 0.68 for citations that lasted 15 days or more. On ChatGPT the link was much weaker.
- Your own site, for facts about you. When a question named a brand, that brand's site stayed cited all 20 days about 28% to 33% of the time, against 4% to 7% for any cited site. In 22 questions about Ahrefs itself, its own pages were cited on 97% to 100% of days for facts like pricing. For reviews and comparisons, Google's AI tools cited Ahrefs' own pages on 5% of days or less, while ChatGPT still cited them often.
Page type made a difference too. In SISTRIX's data, news articles almost never stayed in the stable core (1.4%), while video did best of any type (24%).
If you run a small business, keep your pricing, plans and features pages clear and current, because engines lean on your own site for those facts. For "is it any good" and "X vs Y" questions, Google's AI tools mostly cite other people's pages, so getting named on the review and comparison pages they already cite matters more there.
How should you measure AI visibility if citations keep changing?
Count over time, not once. Ahrefs suggests counting, for each question you track, how many of the last seven days you were cited. A screenshot of one answer proves almost nothing, whether it shows you or not.
SISTRIX's October 2026 study of 10,826 daily measurement series adds a useful split:
- To know how visible you are, ask many different questions. Say a brand really shows up in 25% of answers. Ask one question on 20 days and your measurement could land anywhere from 0% to 59%. Ask 20 different questions once and it lands between 16% and 35%. Ask 100 and it lands between 21% and 29%.
- To know whether something changed, repeat the same questions and compare weekly averages. On a fixed set of 20 questions, one day is accurate to about ±4.8 points and a weekly average to about ±1.8. The daily average typically moves 4 to 5 points, and 10 points or more on one day in ten, with nothing changed at all.
- Track mentions as well as citations. Brand names held up better than links in Ahrefs' data: 50% to 70% of cited websites appeared on only one day, against about one in five brand names. SISTRIX found the same: 20% to 55% of a brand's prompts changed each week, but its total mentions usually moved only 4% to 10%.
- Split brand questions from "best X" questions. On ChatGPT and AI Mode, 64% to 76% of cited pages changed from one day to the next whatever the question type. Questions that named a brand were the steadiest, and "best X" questions were among the least stable, so a strong result on one group can hide a weak one on the other.
That's how Named works. We keep one fixed prompt set and repeat it to check changes, on every AI engine you care about, starting with ChatGPT, Claude, Gemini, Perplexity, Grok and Muse. Each engine gets its own count, because each engine builds its own shortlist. The Namies draft fixes, and a person checks every change. No one can promise an AI will cite you, and results vary run to run.
When should you worry about a lost citation?
When it stays gone for two or three weeks. That's the line Ahrefs draws. Most lost citations come back within days, so don't rewrite a page over one missing citation.
If a citation is still missing after a few weeks, check the page:
- Does it still rank in Google for the question? For AI Overviews and AI Mode, ranking at all mattered more than the exact position.
- Does the title say what the question asks? Rewrite a vague title so it names the question.
- Is the answer up to date? Ahrefs saw a competitor's regularly updated pricing survey overtake its own in AI citations. Ahrefs presents that as an anecdote, not a test.
- Did someone else take the spot? On Google's tools and ChatGPT, a lost citation almost always meant another site took the place, because those answers nearly always cite something. Look at who's cited now.
It works the other way too. If you change a page and the citation comes back the next day, don't credit the edit yet. Citations come back on their own all the time. Look for a steady rise in how often you're cited over several weeks.
FAQ
Why was my site cited by ChatGPT yesterday but not today?
Because ChatGPT picks its sources fresh for each answer. In Ahrefs' study, 59% of the websites ChatGPT cited on day one were gone the next day, and most of the lost ones came back later.
Do AI citations come back after they disappear?
Usually. About 75% of lost citations returned within 20 days in Ahrefs' study, and about half of those returned within three days.
Which AI engine keeps its citations longest?
Google AI Overviews, by most measures. In Ahrefs' data they lost the fewest citations by day two (50%). SISTRIX found no AI Overview source changed over 17 weeks in 53% of prompts, while ChatGPT replaced 74% of its cited domains every week.
How often should I check my AI citations?
Often enough to see a pattern. Ahrefs suggests counting how many of the last seven days you were cited, and SISTRIX suggests comparing weekly averages on a fixed set of questions. A single check tells you very little.
Can anyone guarantee a citation that stays?
No. No one controls what an AI engine cites, and results vary run to run.
Sources
- Ahrefs, "How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations", Mateusz Makosiewicz with Xibeijia Guan, Oct 29, 2026: https://ahrefs.com/blog/how-stable-are-ai-citations/
- SISTRIX, "AI Citation drift: How stable are sources in AI search results?", published May 1, 2026, updated May 12, 2026: https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/
- SISTRIX, "Measuring AI visibility: why breadth beats repetition", published Oct 28, 2026, updated Nov 3, 2026: https://www.sistrix.com/blog/measuring-ai-visibility/
- Mahe Chen and Nick Koudas, "What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)", arXiv:2610.15033, submitted Oct 10, 2026: https://arxiv.org/abs/2610.15033
- Named official facts: https://named.khanex.ai/facts/
Named works to get small businesses recommended by every AI engine you care about, starting with ChatGPT, Claude, Gemini, Perplexity, Grok and Muse. We keep one fixed prompt set and repeat it to check changes. The Namies draft fixes, and a person checks every change. See how Named works.